Driving track determination method, vehicle, electronic equipment and program product

Abstract

The invention provides a driving track determination method, a vehicle, electronic equipment and a program product, and relates to the technical field of intelligent driving. According to the method for determining the driving track, an obtained current candidate driving track set and a current self-vehicle feature vector are input into a track scoring model to score the driving track, so that decoupling of track generation and track evaluation is realized; according to the invention, the driving track scoring is carried out on each driving track in the current candidate driving track set through the track scoring model, the scoring of the driving tracks by adopting an end-to-end deep learning model is realized, the complex logic of an artificial rule base in the related technology is replaced, and the complexity and development and maintenance cost of the intelligent driving system are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and in particular to a driving trajectory determination method, a vehicle, an electronic device and a program product. BACKGROUND

[0002] In intelligent driving such as an Advanced Driver Assistance Systems (ADAS) system and an Autonomous Driving (AD) system, a vehicle needs to generate multiple candidate driving trajectories according to real-time perception information (such as obstacle position, traffic signal, road speed limit, etc.), and evaluate the safety, comfort, efficiency and other indicators of each candidate driving trajectory, and finally select the optimal driving trajectory for execution to cope with complex, dynamic and unpredictable traffic environments.

[0003] In related technologies, the evaluation and screening of candidate driving trajectories are usually implemented by adopting a trajectory generation and trajectory evaluation and screening phased architecture. Specifically, in the driving trajectory generation phase, an environment information (such as obstacle historical trajectory, navigation map, traffic signal, etc.) is received by a deep generation model (such as Transformer, GAN, VAE, etc.), a distribution of a large amount of human driving data is learned, multiple (M) candidate driving trajectories that are statistically feasible, safe and human-like are generated, and a current candidate driving trajectory set is constituted. In the trajectory evaluation and screening phase, for each candidate driving trajectory in the current candidate driving trajectory set, a rule algorithm is used to calculate multiple evaluation indicators (such as safety, comfort, etc.) of each candidate driving trajectory, and further based on a fixed priority order set by the multiple indicators, a pairwise comparison is performed (such as if the safety score of trajectory A is higher than that of trajectory B, then A is selected; if the safety scores are the same, the comfort scores are compared, and so on), and finally an optimal driving trajectory that satisfies all rule constraints is screened out.

[0004] However, in the above driving trajectory evaluation and screening method, all evaluation indicators, priority orders and decision logics depend on artificial pre-setting and debugging, resulting in an extremely large and complex rule base, and problems such as complex and rigid intelligent driving system and high development and maintenance cost. SUMMARY

[0005] The driving trajectory determination method, vehicle, electronic device and program product provided by the embodiments of the present application solve the problem that in related technologies, the evaluation and screening of candidate driving trajectories are usually implemented by adopting a trajectory generation and trajectory evaluation and screening phased architecture, resulting in a complex and rigid intelligent driving system and high development and maintenance cost.

[0006] In a first aspect, the embodiments of the present application provide a driving trajectory determination method, including: obtaining a current candidate driving trajectory set corresponding to a target vehicle and a current ego vehicle feature vector; inputting the current candidate driving trajectory set and the current ego vehicle feature vector into a trajectory scoring model to perform driving trajectory scoring, to obtain a driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model; performing probabilistic processing on the driving trajectory score vector, to obtain a driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set; and determining a target driving trajectory from the current candidate driving trajectory set according to the driving trajectory probability distribution vector.

[0007] In a possible implementation, the trajectory scoring model includes a feature embedding and fusion module, a global enhancement module, and a trajectory scoring module, and inputting the current candidate driving trajectory set and the current ego vehicle feature vector into the trajectory scoring model to perform driving trajectory scoring to obtain the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model includes: inputting the current candidate driving trajectory set and the current ego vehicle feature vector into the feature embedding and fusion module to perform feature embedding and fusion processing, to obtain a first global feature vector output by the feature embedding and fusion module; inputting the first global feature vector into the global enhancement module to perform global context enhancement processing, to obtain a second global feature vector output by the global enhancement module, M being a total number of driving trajectories in the current candidate driving trajectory set; and inputting the second global feature vector into the trajectory scoring module to perform driving trajectory scoring, to obtain the driving trajectory score vector output by the trajectory scoring module.

[0008] In a possible implementation, the feature embedding and fusion module includes a first embedding network and a second embedding network, and inputting the current candidate driving trajectory set and the current ego vehicle feature vector into the feature embedding and fusion module to perform feature embedding and fusion processing to obtain the first global feature vector output by the feature embedding and fusion module includes: inputting the current candidate driving trajectory set into the first embedding network to perform time series encoding, to obtain a trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network; inputting the current ego vehicle feature vector into the second embedding network to perform global feature extraction processing, to obtain a first ego vehicle feature vector output by the second embedding network; broadcasting the first ego vehicle feature vector in M dimensions to obtain a second ego vehicle feature vector; and performing fusion processing on the trajectory feature vector and the second ego vehicle feature vector to obtain the first global feature vector.

[0009] In a possible implementation, the trajectory scoring module includes a fusion embedding network and a classification layer, the second global feature vector is input into the trajectory scoring module for driving trajectory scoring, and a driving trajectory scoring vector output by the trajectory scoring module is obtained, including: inputting the second global feature vector into the fusion embedding network for feature extraction processing, to obtain a third global feature vector output by the fusion embedding network; performing fusion processing on the third global feature vector and the second ego vehicle feature vector, to obtain a fourth global feature vector; inputting the fourth global feature vector into the classification layer for mapping processing, to obtain a driving trajectory scoring vector output by the classification layer.

[0010] In a possible implementation, the trajectory scoring model is obtained by the following method: obtaining training samples, the training samples including a plurality of different driving scenes, and a candidate driving trajectory set and an ego vehicle feature vector corresponding to each driving scene; for each candidate driving trajectory in the candidate driving trajectory set, labeling a true value score for the candidate driving trajectory; inputting the training samples into the pre-trained trajectory scoring model, to obtain a predicted score corresponding to each candidate driving trajectory output by the pre-trained trajectory scoring model; based on a pairwise ranking loss function, according to the predicted score and the true value score, iteratively optimizing the pre-trained trajectory scoring model until the pre-trained trajectory scoring model converges, to obtain the trajectory scoring model.

[0011] In a possible implementation, the true value score is labeled for the candidate driving trajectory, including: calculating a human-likeness score corresponding to the candidate driving trajectory based on a human-likeness rule function; calculating a safety score corresponding to the candidate driving trajectory based on a safety rule function; and generating the true value score according to the human-likeness score and the safety score, and labeling the true value score for the candidate driving trajectory.

[0012] In a possible implementation, the current candidate driving trajectory set and the current ego vehicle feature vector corresponding to the target vehicle are obtained, including: collecting current surrounding environment information and current ego vehicle state information corresponding to the target vehicle; performing feature extraction on the current ego vehicle state information, to obtain the current ego vehicle feature vector; inputting the current surrounding environment information and the current ego vehicle feature vector into the driving trajectory generation model for driving trajectory generation, to obtain the current candidate driving trajectory set output by the driving trajectory generation model.

[0013] In a second aspect, an embodiment of the present application provides a driving trajectory determination apparatus, including:

[0014] The obtaining module is configured to obtain a current candidate driving trajectory set and a current ego vehicle feature vector corresponding to a target vehicle.

[0015] The scoring module is configured to input the current candidate driving trajectory set and the current ego vehicle feature vector into a trajectory scoring model for driving trajectory scoring, to obtain a driving trajectory scoring vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model.

[0016] a processing module configured to probabilize the driving trajectory score vector to obtain a driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set;

[0017] a determining module configured to determine a target driving trajectory from the current candidate driving trajectory set according to the driving trajectory probability distribution vector.

[0018] In a possible implementation, the trajectory scoring model comprises a feature embedding and fusion module, a global enhancement module, and a trajectory scoring module, and the scoring module is specifically configured to: input the current candidate driving trajectory set and the current ego vehicle feature vector into the feature embedding and fusion module for feature embedding and fusion processing to obtain a first global feature vector output by the feature embedding and fusion module; input the first global feature vector into the global enhancement module for global context enhancement processing to obtain a second global feature vector output by the global enhancement module, M being a total number of driving trajectories in the current candidate driving trajectory set; and input the second global feature vector into the trajectory scoring module for driving trajectory scoring to obtain a driving trajectory score vector output by the trajectory scoring module.

[0019] In a possible implementation, the feature embedding and fusion module comprises a first embedding network and a second embedding network, and the scoring module is further configured to: input the current candidate driving trajectory set into the first embedding network for time series encoding to obtain a trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network; input the current ego vehicle feature vector into the second embedding network for global feature extraction processing to obtain a first ego vehicle feature vector output by the second embedding network; broadcast the first ego vehicle feature vector in the M dimension to obtain a second ego vehicle feature vector; and fuse the trajectory feature vector and the second ego vehicle feature vector to obtain the first global feature vector.

[0020] In a possible implementation, the trajectory scoring module comprises a fusion embedding network and a classification layer, and the scoring module is further configured to: input the second global feature vector into the fusion embedding network for feature extraction processing to obtain a third global feature vector output by the fusion embedding network; fuse the third global feature vector and the second ego vehicle feature vector to obtain a fourth global feature vector; and input the fourth global feature vector into the classification layer for mapping processing to obtain the driving trajectory score vector output by the classification layer.

[0021] In one possible implementation, the trajectory scoring model is trained as follows: Training samples are obtained, including multiple different driving scenarios, a set of candidate driving trajectories for each scenario, and a vehicle feature vector; for each candidate driving trajectory in the candidate driving trajectory set, a ground truth score is assigned; the training samples are input into a pre-trained trajectory scoring model to obtain a predicted score for each candidate driving trajectory output by the pre-trained trajectory scoring model; based on a pairwise ranking loss function, the pre-trained trajectory scoring model is iteratively optimized according to the predicted score and the ground truth score until the pre-trained trajectory scoring model converges, thus obtaining the trajectory scoring model.

[0022] In one possible implementation, the driving trajectory determination device further includes a generation module (not shown), which is used to: calculate an anthropomorphic score corresponding to the candidate driving trajectory based on an anthropomorphic rule function; calculate a safety score corresponding to the candidate driving trajectory based on a safety rule function; generate a truth score based on the anthropomorphic score and the safety score, and label the candidate driving trajectory with the truth score.

[0023] In one possible implementation, the acquisition module is specifically used to: collect the current surrounding environment information and the current vehicle status information corresponding to the target vehicle; extract features from the current vehicle status information to obtain the current vehicle feature vector; input the current surrounding environment information and the current vehicle feature vector into the driving trajectory generation model to generate the driving trajectory, and obtain the current candidate driving trajectory set output by the driving trajectory generation model.

[0024] Thirdly, embodiments of this application provide a vehicle, including a vehicle body and an intelligent driving system, the intelligent driving system being used to implement the method for determining the driving trajectory as provided in the first aspect above.

[0025] Fourthly, embodiments of this application provide an electronic device, including:

[0026] At least one processor;

[0027] and memory that is communicatively connected to at least one processor;

[0028] The memory is used to store at least one processor-executable instruction, which is executed by at least one processor to enable at least one processor to perform the method for determining the driving trajectory provided in the first aspect.

[0029] Fifthly, this application also provides a mobile platform that includes electronic devices. Exemplary examples include vehicles, robotic platforms (such as server robots, exploratory robots, scientific research robots, etc.), drones, and other devices, but are not limited thereto.

[0030] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions that are executed by a processor to implement the method for determining a driving trajectory provided in the first aspect.

[0031] In a seventh aspect, embodiments of this application provide a program product comprising computer-executable instructions. When the computer-executable instructions are executed, they implement the method for determining the driving trajectory provided in the first aspect.

[0032] The driving trajectory determination method, vehicle, electronic device, and program product provided in this application embodiment obtain the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle. These are then input into a trajectory scoring model for driving trajectory scoring, resulting in a driving trajectory score vector output by the model corresponding to the current candidate driving trajectory set. Further probabilistic processing is applied to the driving trajectory score vector to obtain a driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set. Finally, the target driving trajectory is determined from the current candidate driving trajectory set based on this probability distribution vector. This application decouples trajectory generation from trajectory evaluation by inputting the obtained current candidate driving trajectory set and current vehicle feature vector into the trajectory scoring model for driving trajectory scoring. The trajectory scoring model scores each driving trajectory in the current candidate driving trajectory set, achieving end-to-end deep learning model scoring of driving trajectories. This replaces the complex logic of manual rule bases in related technologies, reducing the complexity and development / maintenance costs of intelligent driving systems. Attached Figure Description

[0033] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0034] Figure 1 A flowchart illustrating the process of evaluating and filtering driving trajectories;

[0035] Figure 2 Flowchart of the method for determining driving trajectory provided in the embodiments of this application Figure 1 ;

[0036] Figure 3 This is a schematic diagram of the trajectory scoring model provided in the embodiments of this application;

[0037] Figure 4 Flowchart of the method for determining driving trajectory provided in the embodiments of this application Figure 2 ;

[0038] Figure 5A flowchart illustrating the training method of the trajectory scoring model provided in the embodiments of this application;

[0039] Figure 6 A schematic diagram of the structure of the driving trajectory determination device provided in the embodiments of this application;

[0040] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0041] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0042] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0043] In related technologies, a phased architecture of trajectory generation, evaluation, and screening is typically used to evaluate and screen candidate driving trajectories. Specifically, in the trajectory evaluation and screening phase, for each candidate driving trajectory in the current candidate trajectory set, multiple evaluation indicators (such as safety and comfort) are calculated based on rule-based algorithms. Furthermore, pairwise comparisons are performed based on a fixed priority order set for these indicators, ultimately selecting the optimal driving trajectory that satisfies all rule constraints. Figure 1 This is a flowchart illustrating the process of evaluating and filtering driving trajectories. For example... Figure 1 As shown, the implementation method for evaluating and filtering driving trajectories includes the following steps:

[0044] 1) Quantitative Evaluation: Based on preset manual rules, a series of quantifiable evaluation indicators are calculated for each candidate driving trajectory, such as safety (characterized by the minimum safe distance and relative speed between the target vehicle and static obstacles and dynamic traffic participants), comfort (characterized by the curvature, lateral acceleration, and longitudinal acceleration of the driving trajectory), compliance (characterized by the degree of conformity with traffic rules such as lane lines and speed limits), and efficiency (characterized by the deviation from the desired speed).

[0045] 2) Rule ranking: Set fixed priorities based on experience for each evaluation indicator in step 1) (e.g., safety is always the first priority, followed by compliance, and then comfort). The evaluation process of driving trajectory is usually a multi-round pairwise comparison ranking process, following rules such as "if the safety score of trajectory A is higher than that of trajectory B, then select A; if the safety scores are the same, then compare the comfort scores, and so on".

[0046] 3) Filtering decision: Through the layer-by-layer rule filtering and comparison in step 2), the optimal driving trajectory that satisfies all rule constraints is finally selected.

[0047] However, in the aforementioned driving trajectory evaluation and screening methods, all evaluation indicators, priority orders, and decision-making logic rely on manual pre-setting and debugging, resulting in an exceptionally large and complex rule base. When faced with massive amounts of long-tail scenarios, the need to continuously supplement and adjust scenario recognition rules not only leads to extremely high development, testing, and maintenance costs but also makes the system difficult to iterate and expand, resulting in a complex and rigid intelligent driving system with high development and maintenance costs.

[0048] Based on the technical problems existing in related technologies, this application provides a data-driven deep learning model, namely a trajectory scoring model, which replaces the manual rule base. The trajectory scoring model acts as the decision-maker of the entire trajectory planning system, directly receiving a set of candidate driving trajectories and the vehicle's feature vector as input. It scores each driving trajectory in the candidate driving trajectory set and outputs a comprehensive score for each driving trajectory. The trajectory with the highest score is determined as the target driving trajectory. This transforms the complex multi-indicator trade-off problem into a trajectory scoring regression problem, reducing the complexity of the intelligent driving system and the development and maintenance costs.

[0049] The application scenarios of the embodiments of this application will be described below first.

[0050] The driving trajectory determination method provided in this application is applicable to scenarios including but not limited to the following fields: 1) the automotive industry, such as the development of L2 to L4 level autonomous driving systems or assisted driving systems; 2) the field of artificial intelligence and robotics, such as path planning and decision-making systems for mobile robots, unmanned delivery vehicles, and automated guided vehicles (AGVs); 3) the transportation and logistics field, such as vehicle collaborative decision-making and control in intelligent transportation systems. For example, this driving trajectory determination method can be specifically applied to intelligent driving systems of vehicles, decision-making and planning modules in intelligent driving simulation testing platforms, and motion planning packages in first-generation robot operating systems (ROS) or second-generation robot operating systems (ROS2).

[0051] The technical solution of this application and how it solves the above-mentioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0052] Figure 2 Flowchart of the method for determining driving trajectory provided in the embodiments of this application Figure 1 .like Figure 2 As shown, the specific implementation of this method for determining the driving trajectory may include the following steps:

[0053] S201, obtain the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle.

[0054] For example, the target vehicle can be a vehicle that integrates an intelligent driving system such as an autonomous driving system or an assisted driving system.

[0055] For example, the current candidate driving trajectory set includes at least one candidate driving trajectory.

[0056] For example, each candidate driving trajectory in the current candidate driving trajectory set is used to represent the candidate driving trajectory of the target vehicle over a future period of time.

[0057] For example, the current candidate driving trajectory set can be generated by a driving trajectory generation model. This driving trajectory generation model can be a large deep generative network such as a Transformer, a Generative Adversarial Network (GAN), or a Variational Autoencoder (VAE). This application does not limit the specific type of driving trajectory generation model; it can be determined according to the actual application requirements.

[0058] In this step, one possible implementation of the current candidate driving trajectory set is as follows: input the current surrounding environment information (such as historical obstacle trajectories, navigation maps, traffic signals, etc.) corresponding to the target vehicle into the driving trajectory generation model. The driving trajectory generation model learns the distribution of massive human driving data and generates multiple (M) statistically feasible, safe and human-like candidate driving trajectories, which constitute the current candidate driving trajectory set.

[0059] For example, the current vehicle feature vector can be a global feature vector representing the current surrounding environment features and the current vehicle state features corresponding to the target vehicle. The current surrounding environment features include, but are not limited to, static obstacle features and dynamic obstacle features, while the current vehicle state information includes, but is not limited to, the target vehicle's current speed, current acceleration, current navigation features, and the vehicle's current lane.

[0060] For example, in this step, the current vehicle feature vector can be obtained in the following way: the current vehicle state information corresponding to the target vehicle is collected by sensors such as wheel speed sensors and inertial measurement units integrated in the target vehicle, such as the current speed, current acceleration, current position, current orientation and historical speed of the target vehicle; the collected current state information is used to extract features of high-dimensional semantic information through convolutional layers or MLP layers to obtain the current vehicle feature vector corresponding to the target vehicle.

[0061] S202, input the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model to score the driving trajectory, and obtain the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model.

[0062] For example, each candidate driving trajectory in the current candidate driving trajectory set consists of trajectory points of T consecutive time steps, and each trajectory point contains information in at least two dimensions (such as longitudinal displacement y and lateral displacement x).

[0063] For example, the input to the current candidate driving trajectory set can be represented as a vector 1 x M x T x 2 (BatchSize=1). Here, M represents the total number of candidate trajectories in the current candidate driving trajectory set.

[0064] For example, the input of the current vehicle feature vector can be represented as vector 1 x C x 1 x 1. Here, C represents the feature dimension, such as a value of 64, etc.

[0065] In this step, one possible implementation is as follows: input the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model. The trajectory scoring model scores each candidate driving trajectory in the current candidate driving trajectory set based on the current vehicle feature vector, and obtains the driving trajectory score corresponding to each candidate driving trajectory. Finally, the trajectory scoring model outputs the driving trajectory score vector corresponding to the current candidate driving trajectory set.

[0066] For example, the driving trajectory score vector can be represented as 1 x 1 x M x 1.

[0067] S203, perform probabilistic processing on the driving trajectory score vector to obtain the driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set.

[0068] In this step, one possible implementation is to apply an exponential normalization function (softmax function) to the driving trajectory score corresponding to each candidate driving trajectory in the driving trajectory score vector, convert the driving trajectory score corresponding to each candidate driving trajectory into a driving trajectory probability, and obtain the driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set.

[0069] S204, Determine the target driving trajectory from the current candidate driving trajectory set based on the driving trajectory probability distribution vector.

[0070] One possible implementation of this step is to determine the candidate driving trajectory corresponding to the maximum driving trajectory probability in the driving trajectory probability distribution vector as the target driving trajectory.

[0071] For example, the target driving trajectory can be the execution trajectory of the planning system integrated in the target vehicle within the current control cycle. This application embodiment does not limit the current control cycle; it can be determined based on actual application requirements.

[0072] In this embodiment, the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle are obtained. These are then input into a trajectory scoring model for driving trajectory scoring, resulting in a driving trajectory score vector output by the model. This vector is further probabilistically processed to obtain a driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set. The target driving trajectory is then determined from the current candidate driving trajectory set based on this probability distribution vector. This embodiment decouples trajectory generation from trajectory evaluation by inputting the obtained current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model. The model scores each driving trajectory in the current candidate driving trajectory set, employing an end-to-end deep learning model for driving trajectory scoring. This replaces the complex logic of manual rule bases in related technologies, significantly reducing the complexity of the intelligent driving system and the cost and time required for algorithm development, debugging, and subsequent maintenance.

[0073] It is understood that in the driving trajectory determination method provided in the embodiments of this application, when facing a new scenario, it is only necessary to add the corresponding data and retrain the driving trajectory scoring model, without rewriting the rules.

[0074] The following is combined with Figure 3 The trajectory scoring model provided in the embodiments of this application will be described in detail.

[0075] Figure 3 This is a schematic diagram of the trajectory scoring model provided in an embodiment of this application. Figure 3 As shown, the trajectory scoring model includes a first embedding network, a second embedding network, a global enhancement module, a fusion embedding network, and a classification layer (CLS). The first embedding network can also be described as an EgoTrajectoryEmbed network, and the second embedding network can also be described as an EgoFeatureEmbed network.

[0076] For example, the first embedded network may include multiple convolutional layers (Conv1D) and activation functions.

[0077] For example, the second embedded network can be a multilayer perceptron (MLP).

[0078] like Figure 3 As shown, the first embedding network is used to receive the current candidate driving trajectory set, perform time-series encoding on the current candidate driving trajectory set, and output the trajectory feature vector corresponding to the current candidate driving trajectory set;

[0079] The second embedding network is used to receive the current vehicle feature vector, perform global feature extraction processing on the current vehicle feature vector, and output the first vehicle feature vector corresponding to the current vehicle feature vector.

[0080] The global enhancement module is used to receive the first global feature vector, perform global context enhancement processing on the first global feature vector, and output the second global feature vector.

[0081] A fusion embedding network is used to receive the second global feature vector, perform feature fusion processing on the second global feature vector, and output the third global feature vector.

[0082] The classification layer receives the second vehicle feature vector and the third global feature vector, and scores the driving trajectory based on the second vehicle feature vector and the third global feature vector, outputting a driving trajectory score vector.

[0083] From the above Figure 3 As can be seen, the trajectory scoring model provided in this application receives the current candidate driving trajectory set and the current vehicle feature vector, performs forward propagation through its lightweight network, and outputs a driving trajectory scoring vector of length M.

[0084] Compared to related technologies, which require calculating multiple evaluation indicators for each of the M candidate driving trajectories and performing M... 2The large-scale, multi-round pairwise comparisons lead to computational complexity in the evaluation process, consuming significant computing power from onboard computing units. On resource-constrained embedded platforms, this can become a performance bottleneck, resulting in low computational efficiency. This application's embodiment, through highly parallelized forward propagation of the trajectory scoring model, and with a constant computational load after the trajectory scoring model is solidified, significantly improves computational efficiency and better meets the high-efficiency deployment requirements of onboard chips.

[0085] It should be noted that the neural network architecture of the trajectory scoring model provided in this application includes, but is not limited to: on the one hand, adopting dual-stream input and embedding, that is, embedding the current candidate driving trajectory set and the current vehicle feature vector through the first embedding network and the second embedding network respectively; on the other hand, adopting a global enhancement module, performing global average pooling on the trajectory number M dimension of the candidate driving trajectory, and adding it with the first global feature vector, which is the key to realizing global collaborative decision-making; furthermore, performing residual addition on the second vehicle feature vector and the third global feature vector can ensure efficient information flow.

[0086] It is understood that the trajectory scoring model provided in this application has few parameters and high computational efficiency, making it suitable for real-time inference immediately following a large network, thus achieving a balance between performance and efficiency.

[0087] The following is combined with Figure 4 The specific implementation method of step S201, which inputs the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model to perform driving trajectory scoring and obtains the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model, is explained in detail.

[0088] Figure 4 Flowchart of the method for determining driving trajectory provided in the embodiments of this application Figure 2 .like Figure 4 As shown, a specific implementation of the method for determining the driving trajectory, which inputs the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model to score the driving trajectory and obtain the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model, may include the following steps:

[0089] S401, input the current candidate driving trajectory set and the current vehicle feature vector into the feature embedding and fusion module for feature embedding and fusion processing, and obtain the first global feature vector output by the feature embedding and fusion module.

[0090] For example, the feature embedding and fusion module includes, as follows: Figure 3 The first and second embedded networks are shown in the figure.

[0091] Optionally, one possible implementation of this step is as follows: input the current candidate driving trajectory set into the first embedding network for temporal encoding to obtain the trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network; input the current vehicle feature vector into the second embedding network for feature processing to obtain the first vehicle feature vector output by the second embedding network; broadcast the first vehicle feature vector in the M dimension to obtain the second vehicle feature vector; and fuse the trajectory feature vector and the second vehicle feature vector to obtain the first global feature vector.

[0092] The first embedded network is similar to the one described above, and will not be repeated here.

[0093] For example, one possible implementation of inputting the current candidate driving trajectory set into the first embedding network for temporal encoding to obtain the trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network can be: inputting the current candidate driving trajectory set into the first embedding network, and the first embedding network encoding the temporal structure of each candidate driving trajectory in the current candidate driving trajectory set to obtain the trajectory feature vector corresponding to the current candidate driving trajectory set.

[0094] For example, the temporal structure of each candidate driving trajectory can be the lateral trajectory features and longitudinal trajectory features of the candidate driving trajectory at each time t, that is, the combination of lateral and longitudinal spatial positions at each time t.

[0095] For example, the trajectory feature vector can be represented as 1 x F1 x M x 1. Here, F1 represents the embedding dimension, and for example, the value of F1 can be 128.

[0096] Understandably, the first embedding network maps the input 1 x M x T x 2 of the current candidate driving trajectory set into a high-dimensional feature vector 1 x F1 x M x 1. This feature upscaling process can transform the coordinate sequence of each candidate driving trajectory at each trajectory point into a high-dimensional feature rich in temporal context information. This first embedding network can understand the complex characteristics of the driving trajectory, such as its shape and smoothness.

[0097] For example, by sliding multiple convolutional layers (Conv1D) contained in the first embedded network over time, the time dimension T is compressed into the channel dimension F1, thereby enabling the trajectory feature vector to integrate information from the entire time period.

[0098] The second embedded network is similar to the one described above, and will not be repeated here.

[0099] For example, the first vehicle feature vector can be represented as 1 x F2 x 1 x 1. Here, F2 represents the embedding dimension, and for example, the value of F2 can be 64.

[0100] For example, the second vehicle feature vector can be represented as 1 x F2 x M x 1.

[0101] It is understandable that by broadcasting the first vehicle feature vector in the M-dimensional space and copying it M times, the first vehicle feature vector can be fused with the trajectory features corresponding to each candidate driving trajectory. This allows the first vehicle feature vector to be associated with the trajectory features corresponding to each candidate driving trajectory, ensuring that each candidate driving trajectory can be scored in a unified global context.

[0102] For example, one possible way to fuse the trajectory feature vector and the second vehicle feature vector to obtain the first global feature vector is to concatenate the trajectory feature vector and the second vehicle feature vector in the channel dimension to generate a joint feature vector that fuses trajectory details and global context information, i.e., the first global feature vector.

[0103] For example, the first global feature vector can be represented as 1 x F x M x 1. Where F = F1 + F2, and for example, 1 x F x M x 1 can be 1 x 192 x M x 1.

[0104] S402, the first global feature vector is input into the global enhancement module to perform global context enhancement processing in the M dimension, and the second global average feature vector output by the global enhancement module is obtained.

[0105] Where M is the total number of driving trajectories in the current candidate driving trajectory set.

[0106] The global enhancement module is similar to the one described above, and will not be repeated here.

[0107] One possible implementation of this step is as follows: First, the first global feature vector is input into the global enhancement module. The global enhancement module performs global average pooling on the first global feature vector in the M dimension to obtain a 1 x F x 1 x 1 global feature vector. This feature vector represents the global average feature of all candidate driving trajectories in the current candidate driving trajectory set, which contains the overall context information of the current scene. The global feature vector 1 x F x 1 x 1 is broadcast back to 1 x F x M x 1, and then element-wise added to the first global feature vector to obtain the second global average feature vector output by the global enhancement module.

[0108] It is understood that the trajectory scoring model provided in this application is a lightweight network structure with a global perspective. It enables the scoring process to have context awareness by using global average pooling on the candidate driving trajectory dimension (M).

[0109] Understandably, by broadcasting the global feature vector 1 x F x 1 x 1 back to 1 x F x M x 1, and then adding it element-wise with the first global feature vector, the features of each candidate driving trajectory can be given a global perspective. This allows the trajectory to not only consider its own characteristics when scoring driving, but also to perceive the existence of other candidate driving trajectories, thus making a more global and collaborative decision. This avoids selecting routes that are good in themselves but too similar to or conflict with other candidate driving trajectories, which is similar to a lightweight "self-attention" mechanism.

[0110] S403, input the second global feature vector into the trajectory scoring module to perform driving trajectory scoring, and obtain the driving trajectory scoring vector output by the trajectory scoring module.

[0111] For example, the trajectory scoring module includes, as follows: Figure 3 The diagram shows a fusion embedding network and a classification layer.

[0112] Optionally, one possible implementation of this step is as follows: input the second global feature vector into the fusion embedding network for feature extraction processing to obtain the third global feature vector output by the fusion embedding network; fuse the third global feature vector and the second vehicle feature vector to obtain the fourth global feature vector; input the fourth global feature vector into the classification layer for mapping processing to obtain the driving trajectory score vector output by the classification layer.

[0113] For example, a fused embedding network (FuseEmbed) can be composed of 1x1 convolutions.

[0114] For example, one possible implementation of inputting the second global feature vector into the fusion embedding network for feature extraction processing to obtain the third global feature vector output by the fusion embedding network is as follows: input the second global feature vector into the fusion embedding network, and perform nonlinear transformation and feature dimensionality reduction processing on the second global feature vector through the fusion embedding network to remove redundant global information and obtain the third global feature vector output by the fusion embedding network.

[0115] For example, the third global feature vector can be represented as a feature vector of 1 x F2 x M x 1 (e.g., 1 x 64 x M x 1).

[0116] For example, one possible way to fuse the third global feature vector and the second vehicle feature vector to obtain the fourth global feature vector is to add the third global feature vector and the second vehicle feature vector element by element to obtain the fourth global feature vector.

[0117] It is understandable that by adding the third global feature vector and the second vehicle feature vector element by element, a residual connection can be made between the third global feature vector and the second vehicle feature vector. This residual connection (SkipConnection) can ensure that the crucial current vehicle state information is not buried in the deep network, enhance the backpropagation of gradients, and improve the training stability of the trajectory scoring model.

[0118] For example, the classification layer can be a linear layer such as an MLP layer, or it can be a 1x1 convolution. This application does not limit this; the specific method should be determined according to the actual application requirements.

[0119] For example, one possible implementation of inputting the fourth global feature vector into the classification layer for mapping processing to obtain the driving trajectory score vector output by the classification layer is as follows: input the fourth global feature vector into the classification layer, the classification layer maps a scalar score for each candidate driving trajectory based on the fourth global feature vector, and finally outputs the driving trajectory score vector.

[0120] For example, the driving trajectory score vector can be represented as 1 x 1 x M x 1.

[0121] In this embodiment, the current candidate driving trajectory set and the current vehicle feature vector are input into a feature embedding and fusion module for feature embedding and fusion processing to obtain a first global feature vector output by the feature embedding and fusion module. This first global feature vector is then input into a global enhancement module for global average pooling in the M-dimensional dimension to perform global context enhancement processing, resulting in a second global average feature vector output by the global enhancement module. This second global feature vector is further input into a trajectory scoring module for driving trajectory scoring, yielding a driving trajectory score vector output by the trajectory scoring module. Compared to related technologies, where the system's intelligence level depends entirely on the algorithm engineer's understanding and experience of "good driving," its decision-making performance cannot surpass the cognitive limitations of human experts. It struggles to learn the driving skills and game strategies inherent in massive amounts of high-quality human driving data, which are difficult to describe using rules, resulting in a performance ceiling limited by human experience. The trajectory scoring model provided in this application integrates a feature embedding and fusion module, a global enhancement module, and a trajectory scoring module. It can dynamically and implicitly adjust the weights of different evaluation indicators such as safety, comfort, and efficiency according to the context of the input scenario. This enables a more refined, reasonable, and human-like dynamic trade-off while ensuring safety, effectively solving the problem of neglecting one aspect for another in related technologies based on manual rule bases.

[0122] The following is combined with Figure 5 The training method of the trajectory scoring model provided in the embodiments of this application will be described in detail.

[0123] Figure 5 This is a schematic flowchart illustrating the training method of the trajectory scoring model provided in an embodiment of this application. Figure 5 As shown, the training method for this trajectory scoring model may include the following steps:

[0124] S501, Obtain training samples, which include multiple different driving scenarios, as well as candidate driving trajectory sets and vehicle feature vectors corresponding to each driving scenario.

[0125] For example, multiple different driving scenarios can include left turn scenario, right turn scenario, U-turn scenario, lane change scenario, and straight driving scenario, etc.

[0126] In this step, the candidate driving trajectory set corresponding to each driving scenario can be obtained in the following way: For each driving scenario in multiple different driving scenarios, massive amounts of human driving data (such as environmental information and vehicle status information corresponding to human driving) are collected through sensors such as cameras, millimeter-wave radar, and lidar integrated on the vehicle. The collected human driving data is then input into a pre-trained driving trajectory generation model for model training until the model parameters of the pre-trained driving trajectory converge, thus obtaining the driving trajectory generation model. The learning objective of this model training process is to be able to generate diverse driving trajectories similar to those of humans based on environmental input.

[0127] The driving trajectory generation model is similar to that described above and will not be repeated here. This application does not limit the loss function used when training the pre-trained driving trajectory generation model; it can be determined based on the specific application requirements.

[0128] In one possible implementation, the candidate driving trajectory set corresponding to each driving scenario can be generated by a driving trajectory generation model; in another possible implementation, the candidate driving trajectory set corresponding to each driving scenario can be generated by simulation tools.

[0129] For example, the method for obtaining the vehicle feature vector is similar to that described above, and will not be repeated here.

[0130] S502 assigns a true value score to each candidate driving trajectory in the candidate driving trajectory set.

[0131] In this step, one possible implementation involves calculating a comprehensive quality score as a truth score for each candidate driving trajectory in the candidate driving trajectory set. This truth score is composed of a combination of quantifiable anthropomorphic and safety rule functions.

[0132] Optionally, one possible implementation for labeling candidate driving trajectories with truth scores is as follows: calculate the anthropomorphic score corresponding to the candidate driving trajectory based on the anthropomorphic rule function; calculate the safety score corresponding to the candidate driving trajectory based on the safety rule function; generate truth scores based on the anthropomorphic score and the safety score, and label the candidate driving trajectory with truth scores.

[0133] It should be noted that the anthropomorphism rule function focuses on the horizontal and vertical similarity between the driving trajectory generated by the driving trajectory generation model and the human driving trajectory. Under the premise of high-quality training data, in most cases, the most anthropomorphic driving trajectory is often the best overall driving trajectory, that is, it better balances multiple factors such as safety, standardization, efficiency, and comfort. Safety, as the most basic and important evaluation indicator of driving behavior, is independent of anthropomorphism and is scored separately. The safety rule function focuses on the safe distance between the vehicle (target vehicle) and surrounding dynamic and static obstacles (such as other vehicles, pedestrians, bicycles / electric vehicles, curbs, flower beds, construction areas, etc.). This safe distance is calculated based on the relative position, speed, angle and other information of the current and historical frames, taking into account both safety (whether a collision will occur) and peace of mind (maintaining sufficient spatiotemporal margin in case of collision).

[0134] It is understandable that the computation rules of the aforementioned anthropomorphic rule function and security rule function are only used to generate training labels. Their complexity is far lower than that of online rule systems that rely on various scene recognition logics, and ultimately only provide a scalar value for supervised learning.

[0135] S503: Input the training samples into the pre-trained trajectory scoring model to obtain the predicted score for each candidate driving trajectory output by the pre-trained trajectory scoring model.

[0136] S504, based on the pairwise ranking loss function, iteratively optimizes the pre-trained trajectory scoring model according to the predicted score and the true score until the pre-trained trajectory scoring model converges, thus obtaining the trajectory scoring model.

[0137] For example, the Pairwise Ranking Loss (PRL) function can directly optimize the ranking ability of a pre-trained trajectory scoring model, forcing the pre-trained trajectory scoring model to learn to give a higher score to a better driving trajectory than a second-best driving trajectory in the same scenario.

[0138] It is understood that the trajectory scoring model provided in this application uses a method based on human driving data and rule-generated trajectory quality score labels (GT) to train the trajectory scoring model.

[0139] Compared to related technologies, rule-based methods rely on fixed, either-or priority rankings, lacking a flexible understanding of the scene context. This results in a lack of human-like, flexible decision-making capabilities when facing complex scenarios requiring dynamic trade-offs (e.g., needing to slightly sacrifice comfort or efficiency to avoid a suddenly intervening vehicle). Rule-based systems tend to make rigid, inflexible, or even overly conservative decisions, leading to a poor driving experience. In this application's embodiment, the model is trained using massive amounts of high-quality human driving data, directly learning the decision-making patterns of human drivers under complex multi-objective trade-offs. This makes the driving trajectory selection more human-like and intelligent, significantly improving the driving experience and avoiding the rigidity and robotic feel of rule-based system decisions.

[0140] It is understood that, in the embodiments of this application, a complete offline training method for the trajectory scoring model is proposed, including generating a candidate driving trajectory set for model training using a driving trajectory generation model, generating high-quality ground truth scoring labels using rule functions (safety rule functions and anthropomorphic rule functions), and using a strategy of training with pairwise loss.

[0141] Optionally, one possible implementation of step S201, which involves obtaining the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle, is as follows: collecting the current surrounding environment information and the current vehicle state information corresponding to the target vehicle; extracting features from the current vehicle state information to obtain the current vehicle feature vector; and inputting the current surrounding environment information and the current vehicle feature vector into the driving trajectory generation model to generate the driving trajectory, thereby obtaining the current candidate driving trajectory set output by the driving trajectory generation model.

[0142] The specific implementation method is similar to that described above, and will not be repeated here.

[0143] In summary, it can be understood that the embodiments of this application, as a dedicated evaluation module following the driving trajectory generation model, collaborate with the driving trajectory generation model to form a complete "generation-evaluation" planning paradigm, replacing the traditional "generation-rule evaluation" paradigm. On one hand, a deep learning model, namely a trajectory scoring model, is used to score multiple candidate driving trajectories, obtaining the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model, and the optimal driving trajectory, i.e., the target driving trajectory, is selected based on the driving trajectory score vector. On the other hand, the driving trajectory planning is divided into a two-level architecture of "driving trajectory generation model" + "dedicated evaluation model", and a decision framework of "dual score (human-likeness + safety) real-time dynamic weighting".

[0144] In summary, the method for determining the driving trajectory provided in this application has the following beneficial effects:

[0145] 1) Enhance the human-likeness and smoothness of decision-making: The trajectory scoring model is trained with massive amounts of high-quality human driving data, directly learning the decision-making patterns of human drivers under complex multi-objective trade-offs, making the driving trajectory selection more human-like, smooth, and intelligent, significantly improving the driving experience and avoiding the rigidity and machine-like feel of rule-based system decision-making.

[0146] 2) Significantly reduce development and maintenance complexity: By adopting a single end-to-end learning model instead of a large and complex manual rule base, the cost and cycle of algorithm development, debugging and later maintenance are greatly reduced. When facing new scenarios, only the corresponding data needs to be added and the model needs to be retrained, without rewriting the rules.

[0147] 3) Achieve more accurate dynamic trade-offs: The internal mechanism of the trajectory scoring model can dynamically and implicitly adjust the weights of different evaluation indicators (safety, comfort, efficiency, etc.) according to the context of the input scenario. Under the premise of ensuring safety, it can make more refined, reasonable and human decision-making logic dynamic trade-offs, solving the problem of rule systems neglecting one aspect for another.

[0148] 4) Improved computational efficiency: The trajectory scoring model features highly parallel forward propagation, and the computational load remains constant after the trajectory scoring model is solidified. Compared with rule-based methods that require multiple rounds of iterative comparisons, it can significantly improve computational efficiency and is more suitable for the efficient deployment requirements of automotive chips.

[0149] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0150] Figure 6 This is a schematic diagram of the driving trajectory determination device provided in an embodiment of this application. Figure 6 As shown, the driving trajectory determination device 60 includes an acquisition module 610, a scoring module 620, a processing module 630, and a determination module 640.

[0151] Among them, the acquisition module 610 is used to acquire the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle;

[0152] The scoring module 620 is used to input the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model to score the driving trajectory, and obtain the driving trajectory score vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model.

[0153] The processing module 630 is used to perform probabilistic processing on the driving trajectory score vector to obtain the driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set.

[0154] The determination module 640 is used to determine the target driving trajectory from the current candidate driving trajectory set based on the driving trajectory probability distribution vector.

[0155] In one possible implementation, the trajectory scoring model includes a feature embedding and fusion module, a global enhancement module, and a trajectory scoring module. The scoring module 620 is specifically used to: input the current candidate driving trajectory set and the current vehicle feature vector into the feature embedding and fusion module for feature embedding and fusion processing to obtain a first global feature vector output by the feature embedding and fusion module; input the first global feature vector into the global enhancement module for global context enhancement processing to obtain a second global feature vector output by the global enhancement module, where M is the total number of driving trajectories of all candidate driving trajectories in the current candidate driving trajectory set; and input the second global feature vector into the trajectory scoring module for driving trajectory scoring to obtain a driving trajectory scoring vector output by the trajectory scoring module.

[0156] In one possible implementation, the feature embedding and fusion module includes a first embedding network and a second embedding network. The scoring module 620 is further configured to: input the current candidate driving trajectory set into the first embedding network for temporal encoding to obtain the trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network; input the current vehicle feature vector into the second embedding network for global feature extraction processing to obtain the first vehicle feature vector output by the second embedding network; broadcast the first vehicle feature vector in the M-dimensional space to obtain the second vehicle feature vector; and fuse the trajectory feature vector and the second vehicle feature vector to obtain the first global feature vector.

[0157] In one possible implementation, the trajectory scoring module includes a fusion embedding network and a classification layer. The scoring module 620 is further configured to: input the second global feature vector into the fusion embedding network for feature extraction processing to obtain the third global feature vector output by the fusion embedding network; fuse the third global feature vector and the second vehicle feature vector to obtain the fourth global feature vector; and input the fourth global feature vector into the classification layer for mapping processing to obtain the driving trajectory scoring vector output by the classification layer.

[0158] In one possible implementation, the trajectory scoring model is trained as follows: Training samples are obtained, including multiple different driving scenarios, a set of candidate driving trajectories for each scenario, and a vehicle feature vector; for each candidate driving trajectory in the candidate driving trajectory set, a ground truth score is assigned; the training samples are input into a pre-trained trajectory scoring model to obtain a predicted score for each candidate driving trajectory output by the pre-trained trajectory scoring model; based on a pairwise ranking loss function, the pre-trained trajectory scoring model is iteratively optimized according to the predicted score and the ground truth score until the pre-trained trajectory scoring model converges, thus obtaining the trajectory scoring model.

[0159] In one possible implementation, the driving trajectory determination device further includes a generation module (not shown), which is used to: calculate an anthropomorphic score corresponding to the candidate driving trajectory based on an anthropomorphic rule function; calculate a safety score corresponding to the candidate driving trajectory based on a safety rule function; generate a truth score based on the anthropomorphic score and the safety score, and label the candidate driving trajectory with the truth score.

[0160] In one possible implementation, the acquisition module 610 is specifically used to: collect the current surrounding environment information and the current vehicle status information corresponding to the target vehicle; extract features from the current vehicle status information to obtain the current vehicle feature vector; input the current surrounding environment information and the current vehicle feature vector into the driving trajectory generation model to generate the driving trajectory, and obtain the current candidate driving trajectory set output by the driving trajectory generation model.

[0161] The driving trajectory determination device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0162] This application also provides a vehicle, including: a vehicle body and a planning system, the planning system being used to execute the driving trajectory determination method provided in the above embodiments.

[0163] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 70 provided in this embodiment includes at least one processor 701 and a memory 702. Optionally, the electronic device 70 further includes a communication component 703. The processor 701, memory 702, and communication component 703 are connected via a bus 704.

[0164] In a specific implementation, at least one processor 701 executes computer execution instructions stored in memory 702, causing at least one processor 701 to perform the above-described method.

[0165] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0166] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0167] The memory may include random access memory (RAM) and non-volatile memory (NVM), such as at least one disk storage device.

[0168] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0169] This application also provides a mobile platform, which includes electronic devices. Exemplary examples include vehicles, robotic platforms (such as server robots, exploratory robots, scientific research robots, etc.), drones, and other devices, but are not limited thereto.

[0170] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0171] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0172] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0173] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0174] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0175] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0177] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0178] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0179] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for determining a driving trajectory, characterized in that, include: Obtain the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle; The current candidate driving trajectory set and the current vehicle feature vector are input into the trajectory scoring model to score the driving trajectory, and the driving trajectory score vector corresponding to the current candidate driving trajectory set is output by the trajectory scoring model. The driving trajectory score vector is probabilistically processed to obtain the driving trajectory probability distribution vector corresponding to the current candidate driving trajectory set; The target driving trajectory is determined from the current candidate driving trajectory set based on the driving trajectory probability distribution vector.

2. The method for determining the driving trajectory according to claim 1, characterized in that, The trajectory scoring model includes a feature embedding and fusion module, a global enhancement module, and a trajectory scoring module. The step of inputting the current candidate driving trajectory set and the current vehicle feature vector into the trajectory scoring model to perform driving trajectory scoring, and obtaining the driving trajectory scoring vector corresponding to the current candidate driving trajectory set output by the trajectory scoring model, includes: The current candidate driving trajectory set and the current vehicle feature vector are input into the feature embedding and fusion module for feature embedding and fusion processing to obtain the first global feature vector output by the feature embedding and fusion module; The first global feature vector is input into the global enhancement module for global context enhancement processing to obtain the second global feature vector output by the global enhancement module, where M is the total number of driving trajectories of all candidate driving trajectories in the current candidate driving trajectory set; the second global feature vector is input into the trajectory scoring module for driving trajectory scoring to obtain the driving trajectory scoring vector output by the trajectory scoring module.

3. The method for determining the driving trajectory according to claim 2, characterized in that, The feature embedding and fusion module includes a first embedding network and a second embedding network. The step of inputting the current candidate driving trajectory set and the current vehicle feature vector into the feature embedding and fusion module for feature embedding and fusion processing, to obtain the first global feature vector output by the feature embedding and fusion module, includes: The current candidate driving trajectory set is input into the first embedding network for temporal encoding to obtain the trajectory feature vector corresponding to the current candidate driving trajectory set output by the first embedding network; The current vehicle feature vector is input into the second embedding network for global feature extraction processing to obtain the first vehicle feature vector output by the second embedding network. The first vehicle feature vector is broadcast along the M-dimensional plane to obtain the second vehicle feature vector. The trajectory feature vector and the second vehicle feature vector are fused to obtain the first global feature vector.

4. The method for determining the driving trajectory according to claim 3, characterized in that, The trajectory scoring module includes a fusion embedding network and a classification layer. The step of inputting the second global feature vector into the trajectory scoring module to perform driving trajectory scoring, and obtaining the driving trajectory scoring vector output by the trajectory scoring module, includes: The second global feature vector is input into the fusion embedding network for feature extraction to obtain the third global feature vector output by the fusion embedding network; the third global feature vector and the second vehicle feature vector are fused to obtain the fourth global feature vector. The fourth global feature vector is input into the classification layer for mapping processing to obtain the driving trajectory score vector output by the classification layer.

5. The method for determining a driving trajectory according to any one of claims 1 to 4, characterized in that, The trajectory scoring model was trained in the following way: Acquire training samples, which include multiple different driving scenarios, as well as a set of candidate driving trajectories and a vehicle feature vector corresponding to each driving scenario; For each candidate driving trajectory in the candidate driving trajectory set, a true value score is assigned to the candidate driving trajectory. The training samples are input into the pre-trained trajectory scoring model to obtain the predicted score for each candidate driving trajectory output by the pre-trained trajectory scoring model. Based on the pairwise ranking loss function, the pre-trained trajectory scoring model is iteratively optimized according to the predicted score and the ground truth score until the pre-trained trajectory scoring model converges, thus obtaining the trajectory scoring model.

6. The method for determining a driving trajectory according to claim 5, characterized in that, The step of assigning a ground truth score to the candidate driving trajectory includes: Based on the anthropomorphic rule function, the anthropomorphic score corresponding to the candidate driving trajectory is calculated; Based on the safety rule function, the safety score corresponding to the candidate driving trajectory is calculated; Based on the anthropomorphic score and the safety score, the truth score is generated, and the truth score is labeled for the candidate driving trajectory.

7. The method for determining a driving trajectory according to any one of claims 1 to 4, characterized in that, The step of obtaining the current candidate driving trajectory set and the current vehicle feature vector corresponding to the target vehicle includes: Collect the current surrounding environment information and current vehicle status information corresponding to the target vehicle; Feature extraction is performed on the current vehicle state information to obtain the current vehicle feature vector; The current surrounding environment information and the current vehicle feature vector are input into the driving trajectory generation model to generate a driving trajectory, and the current candidate driving trajectory set output by the driving trajectory generation model is obtained.

8. A vehicle, characterized in that, It includes a vehicle body and a trajectory planning system, the trajectory planning system being used to implement the method for determining a driving trajectory as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method for determining the driving trajectory as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, include: A computer program, when executed by a processor, implements the method for determining a driving trajectory as described in any one of claims 1 to 7.

Images